Re: [PATCH 0/4] x86_64: Optimize percpu accesses

[Jeremy Fitzhardinge <jeremy@goop.org>]

Mike Travis wrote:
> This patchset provides the following:
>
>   * x86_64: Cleanup setup_percpu by fixing some minor potential
>     problems as well as add some debugging aids.
>
>   * x86_64: Rebase per cpu variables to zero
>
>     Rebase per cpu variables to zero in preparation for the following
>     patch to fold the pda into the per cpu area.
>
>   * x86_64: Fold pda into per cpu area
>
>     Declare the pda as a per cpu variable. This will allow the per cpu
>     variables to be accessible on the x86_64 using %gs as the base of
>     the percpu areas for each cpu:
>
> 	%gs:per_cpu_xxxx
>
>   * x86_64: Reference zero-based percpu variables offset from gs
>
>     Actually implement the above operation for __get_cpu_var() and
>     __put_cpu_var().  Since this is now a single instruction, we
>     can remove the non-preemptible versions of x86_read_percpu()
>     and x86_write_percpu().
>   

No, I think you've misunderstood these calls.

get_cpu_var(x) evaluates to an lvalue of this cpu's 'x'.  It disables 
preemption, in the same manner as get_cpu().

put_cpu_var(x) does nothing more than re-enable preemption, to pair with 
get_cpu_var().

__get_cpu_var(x) is the same as get_cpu_var, but it assumes that 
preemption is already disabled.  There is no __put_cpu_var().

The important point is that an expression like "__get_cpu_var(x) = foo" 
does not evaluate to a single instruction, and is not preempt or 
interrupt -atomic.  That's the reason x86_X_percpu() exist, since 
they're a single instruction in an asm.  However, with %gs: based 
addressing they can be easily unified.

    J